1 Introduction
Recent advances in communication and computing technologies have created a paradigm shift to fog radio access networks (RANs) in 5G. Fog RAN enables lowlatency RAN operations at the network edge, by making use of computing and storing capabilities in wirelessly connected edge devices. Beyond 5G, witnessing recent breakthroughs in machine learning (ML), we envisage that these edge devices will have ondevice ML capabilities, thereby locally and promptly carrying out missioncritical RAN operations even when network connectivity is temporarily lost.
There exist two prerequisites for this. One is to utilize usergenerated data at the network edge. Edge devices can observe the freshest surrounding environments, so their private datasets are indispensable for missioncritical decisions. The other one is to train ondevice ML models collectively across devices over wireless links. Every single device observes a small portion of the global environment, mandating the collaborative training without exchanging privacysensitive raw data. These two essential factors call for a novel fog computing based distributed ML framework that is suited for wireless connectivity while preserving local data privacy at the edge, hereafter termed fog ML (FML) [1].
A notable attempt to enable FML is the federated averaging algorithm [2], also known as Vanilla federated learning (FL). At regular intervals, devices in Vanilla FL exchange model parameters, such as weights and gradients, thereby training their local models collectively without revealing private data. While Vanilla FL preserves privacy, it is still far from reflecting RAN characteristics including channel dynamics and network architectures.
Scope and Organization: Towards enabling ML aided fog RAN, this article aims at presenting communicationefficient FML frameworks with high accuracy and data privacy guarantees. To this end, taking Vanilla FL as our baseline, we address the following fundamental questions. [colback=white, colframe=blue!30!black, boxrule=.5pt, boxsep=3pt,left=1.5pt,right=1.5pt,top=2pt,bottom=2pt, sharp corners]
Q1. (Adaptivity) How often do devices need to communicate for given channel conditions and training dynamics?
Q2. (Scalability) How to support a large number of federating devices with limited transmission power?
Q3. (Robustness) How to make FML robust against malfunctions and adversarial attacks? See S1S3 in Sec. 2
Q4. (Large Model) How to deal with largesized models such as
deep neural networks
?Q5. (Asymmetric Channel) How to cope with asymmetric uplink and downlink channel capacities? S4S5 in Sec. 3
Q6. (NonIID Data) How to rectify datasets that are not independent and identically distributed (NonIID)?
Q7. (OneShot Training) How to train FML models in a single communication round? S6S7 in Sec. 4
As shown in Fig. 1, our proposed solutions S1S7 to these problems Q1Q7 boil down to three types of ondevice ML model training methods: by exchanging (1) model parameters, (2) model outputs, and (3) surrogate data, which are elaborated respectively in Sec. 2, 3, and 4, followed by conclusion and discussions about extending our proposed FML frameworks in Sec. 5.
Related Works: For a comprehensive understanding of FML frameworks, readers are encouraged to check [1, 3, 4]. Theoretical and algorithmic enablers of ondevice ML are overviewed in [1]. Privacy preserving aspects of FML frameworks are studied in [3, 4]. Note that [3] provides an FML taxonomy based on data and feature correlations, whereas this article describes FML frameworks from the perspective of what to exchange across devices. It is also noted that this work focuses on communicationefficient dataparallel architectures wherein devices have an identical model structure but different local data. Modelparallel architectures under which a single model is split across devices are explored in [4], ignoring communication efficiency.
2 Model Parameter Exchange
Vanilla FL considers a fixed model averaging interval, and thus cannot cope with instantaneous channel conditions and training dynamics (see Q1). Another drawback is its reliance on a single edge server. The server may not be reachable from faraway devices with low transmission power, due to limited energy or interference mitigation (Q2). Besides, if the server fails due to malfunctions or adversarial attacks, the entire operations will be disrupted, i.e., single point of failure (Q3). To address these problems, this section introduces an advanced FL method adaptive to model training and channel dynamics (S1), as well as serverless FML frameworks with minimum transmission power (S2) and in combination with blockchain (S3).
2.1 Federated Learning with Adaptive Scheduling
In Vanilla FL, each device uploads its local model parameters at fixed communication intervals to an edge server, i.e., parameter server. By averaging the uploaded parameters, a global model is constructed at the server and downloaded by each device for its local model updates, thereby ensuring robustness against locally unseen data. The communication interval governs not only communication overhead but also accuracy of Vanilla FL. Indeed, reducing communication intervals increases communication overhead, while achieving higher accuracy by better synchronizing local models and avoiding overfitting towards local datasets.
To balance the communication overhead and accuracy, (S1) Adaptive FL optimizes its communication interval at the edge server [5]
, so as to minimize an ML loss function subject to limited communication and computation budgets. It is found that the optimal communication interval depends not only on the communication and computation costs, but also on the data distribution and model characteristics that affect the convexity and smoothness of the loss function. The main building blocks of Adaptive FL are described next.
1)
Data Distribution Estimation
: The communication interval should be smaller if the local datasets at different devices are more diverse. To measure this dissimilarity, Adaptive FL evaluates the gradient divergence that quantifies the distance between the gradient computed on a local dataset and the average of all such local gradients. A larger gradient divergence indicates more diverse datasets. The gradient divergence can easily be computed in a distributed manner with negligible computation cost, since the model training algorithms rely fundamentally on distributed gradient descent methods.2) Model Characteristics Estimation: Training an ML model is achieved by minimizing a loss function defined on the data samples and model parameters. The model characteristics can therefore be abstracted into the loss function. In view of this, Adaptive FL estimates the loss function’s Lipschitzness and smoothness parameters in a distributed fashion, thereby capturing the model properties for the communication interval optimization.
3) Cost Measurement: Adaptive FL measures the cost per computing iteration and the cost per communication round, thereby examining whether the total communication and computation budgets are reached. The cost measurements are also used in finding the best communication interval. As an example, for a significant communication cost, a larger communication interval is preferred.
Accounting for these three aspects, a closedform training convergence bound is derived in [5]. This bound plays a key role in optimizing the communication interval of Adaptive FL in a distributed way with low computational complexity. Fig. 2 compares the test accuracy results under Vanilla FL and Adaptive FL with different data distributions, in which the communication and computation costs are given by realworld measurements in [5]. We see that a fixed communication interval as used in Vanilla FL does not work well for all the considered data distributions, because it does not adapt to the data distribution, learning dynamics, and/or cost budget. Furthermore, Fig. 2 shows that the communication interval of Adaptive FL is close to the best achievable interval.
In addition to adjusting the averaging interval in Adaptive FL, communication payload sizes can jointly be optimized under the communication and accuracy tradeoff [1]. To this end, one can adaptively quantize the model parameters to be exchanged by adjusting their arithmetic precision, or exchanging only a fraction of the model parameters, i.e., sparse parameter averaging that is known to achieve the same training convergence rate as periodic averaging [8].
2.2 ServerLess FML
Vanilla FL relies on a single edge server collecting and averaging the model parameters of devices (see Fig. 3a). This serverbased architecture is not scalable, as the maximum transmission power of each device is restricted by limited energy and/or interdevice interference. To overcome the limitation, we revisit the optimization problem of Vanilla FL, and present a serverless FML framework as follows.
In brief, Vanilla FL minimizes each device’s local loss function, and globally averages their local models, thereby yielding a global model at the server. This can be recast by an average consensus problem that minimizes the loss function averaged across devices, i.e., empirical risk, subject to the constraint making each local model equal to the global model. A primaldual decomposition method can effectively solve this problem in a decentralized way, yet is not scalable since the constraint entails exponentially increasing communication rounds with the number of federating devices.
Alternatively, one can consider a relaxed constraint making each local model equal only to its two neighboring devices’ local models. (S2) Group ADMM (GADMM) [6] solves such a modified constrained optimization problem, using the Alternating Direct Method of Multiplier (ADMM) algorithm with a novel device grouping method (see Fig. 3b). GADMM is scalable in a way that it requires each device to communicate only with two neighboring devices, regardless of the total number of federating devices. The operations of GADMM are detailed next.
1) HeadTail Grouping: Devices are divided into two groups head and tail, such that each head device is connected to other devices through its two neighboring tail devices. The devices in the same group update their model parameters in parallel, while the devices in different groups update the model parameters in an alternating way.
2) Primal Updates: At first, each head device locally updates its model parameters, i.e., primal variables. This is done by minimizing the head device’s augmented Lagrange function that depends not on other head devices but on its neighboring tail devices’ previous primal variables. Then, the updated primal variables are broadcasted to the neighboring tail devices. Next, each tail device updates its primal variables in parallel, based on its neighboring head devices’ updated primal variables. Likewise, the updated primal variables are broadcasted to the neighboring head devices. Consequently, at any communication round, only half of the devices broadcast their primal variables.
3) Dual Updates: Finally, each device locally updates the dual variables, based on its current primal variables and the primal variables broadcasted from the neighboring devices. Consequently, GADMM achieves the training convergence rate with iterations, which is faster than of the distributed gradient descent algorithm that commonly forms the basis of Vanilla FL and its variants.
Next, serverless architectures are operated by multiple model parameter aggregating devices. This is useful for avoiding the single point of failure problem, yet makes it challenging to reach a legitimate consensus on a global model. (S3) Blockchained FL (BlockFL) resolves this pressing issue by integrating blockchain into FL [7]. In BlockFL (see Fig. 3c), randomly selected devices, i.e., miners, verify and collect the model parameters of the other devices. Then, by solving a random puzzle, i.e., proofofwork (PoW), the global model is determined by the first winning miner’s aggregated model parameters. Furthermore, by leveraging blockchain, BlockFL provides rewards to devices proportionally to their local training data amount, thereby promoting more federation of devices having larger training data. In general, serverless FML relies solely on low ondevice energy and memory, and trading off its accuracy and robustness against communication efficiency subject to each network topology calls for further investigation.
3 Model Output Exchange
Model parameter exchange as done in Vanilla FL is illsuited for largesized ML models consisting of millions of parameters that consume dozens of gigabytes (see Q4). The problem is aggravated in the uplink whose channel capacity is much lower than downlink capacity (Q5), due to the limited transmission power of devices. Alternatively, this section introduces a model output exchanging FML framework (S4), referred to as federated distillation (FD) whose communication payload size is independent of the model size [9]. In combination with FL in the downlink, exploiting FD in the uplink is effective in coping with asymmetric uplinkdownlink channels (S5).
3.1 Federated Distillation
In traditional ML, largesized models are often handled via pruning and/or reducing the arithmetic precision of model parameters [1]. These model compression and quantization methods are mostly performed after training, or otherwise achieved at the cost of compromising accuracy. In contrast, we present (S4) FD that reduces the communication cost by exchanging model outputs rather than model parameters, without altering local model structures (see Fig. 4).
At its core, FD builds on knowledge distillation (KD) that was originally designed for transferring a pretrained teacher model’s knowledge on its prediction capability to an untrained student model [10]. To this end, KD measures the gap between both models’ output distributions, i.e., logits, of a common data sample, and reflects this gap in updating the student model. Specifically, both teacher and student models in KD first observe a common data sample . The student’s prediction on
is more likely to be wrong, if the student’s logit
of is less similar to the teacher’s logit of . In this case, the student increases a penalty in its loss function, i.e., distillation regularizer, thereby correcting the student’s bias on . FD exploits the said KD operations for FML, while overcoming the following two limitations of KD: (1) pretraining of the teacher model and (2) common sample observations of teacher and student models.FD obviates the problem (1) by adopting codistillation (CD) [10]. The core idea is to make every device become a student, while treating the global average logit of , averaged across devices, as the teacher’s logit of . However, there still exists the requirement (2). It makes CD exchange the local logits as many as the number of data samples , incurring huge communication overhead while leaking private information. To resolve this, FD exchanges not local logits of raw samples, but local average logits of proxy samples , as detailed next.
1)
FD in Supervised Learning
: There are four steps in FD operations. Firstly, each device constructs proxy samples fewer than raw samples. In classification tasks, raw samples can be mapped into their groundtruth labels, and one representative raw sample of the groundtruth label is set as a proxy sample [9]. In Fig. 4, for blue and yellow groundtruth labels, two proxy samples and are thereby constructed. Secondly, during local training iterations, each device stores all the logits originated from each groundtruth label separately, and generates local average logits and at intervals, which are uploaded to the server. Thirdly, by separately averaging the uploaded and across devices, the server produces the global average logits and , which are downloaded by devices. Finally, with proxy samples and , every device runs CD by comparing its own logits and with the downloaded and , respectively, thereby collectively training ondevice ML models.2) FD in Reinforcement Learning
: In reinforcement learning, quantized states can be proxy states
[11]. In the CartPole game, for instance, raw states of angles in can be mapped into two quantized angles , thereby constructing proxy states and . As done in the supervised learning example 1), each device exchanges local average logits associated with proxy states, referred to as proxy experience memory, thereby collectively train ondevice ML models. This reduces communication overhead while preserving more privacy, compared to classical distributed reinforcement learning that exchanges raw experience memory containing all logits associated with raw states. A more general approach to constructing proxy samples will be elaborated in Sec. 4.1.In these ways, FD yields comparable accuracy with Vanilla FL, while reducing the communication overhead by tenfold, in our experiments with reinforcement learning for completing the CartPole game [11]
and supervised learning for classifying the MNIST dataset
[9]. The effectiveness of FD may however be limited when local datasets are nonIID across devices. For example, if one device in Fig. 4 only has blue samples while the other device’s samples are all yellow, then they cannot transfer any knowledge to each other. A more finegrained split of proxy states partly resolves this problem, yet may incur more privacy leakage. Therefore, it is essential in FD to rectify nonIID datasets, which will be discussed in Sec. 4.3.2 Federated Learning after Distillation
By marrying FD and FL, we propose (S5) federated learning after distillation (FLD). In FLD (see Fig. 5a), the uplink and downlink operations follow FD and Vanilla FL, respectively. This necessitates a conversion from the uploaded logits to the model parameters to be downloaded. The said conversion is enabled by performing KD at the server lying in between the uplink and downlink. There are two design motivations of FLD as follows.
1) UplinkDownlink Capacity Asymmetry: Uplink capacity is often limited by the low transmission power budget of each device [12]. This may cause a bottleneck in Vanilla FL particularly with largesized models. In the uplink, it is therefore suitable to exchange logits as done in FD. By contrast, in the downlink with much higher capacity, exchanging model parameters as done in Vanilla FL is preferable. This choice is based on our observations in the MNIST classification experiments [9], in which FL yields slightly higher accuracy than FD, whereas FD consumes 10x smaller bandwidth.
2) Data Distribution Reflecting Aggregation: FD constructs its global average logits for the downlink, by averaging the local average logits across devices. A simple average cannot account for the data distribution over devices. A weighted average can partly ameliorate this issue, yet the optimal weight selection brings about another challenging problem. By contrast, FLD forms its global model parameters by converting the global average logits to a global model, via KD with few seed samples uploaded from devices. These seed samples are not only required for running KD, but also useful for taking into account the data distribution over devices.
The performance of FLD hinges on collecting seed samples in a communicationefficient way while preserving data privacy. One simple solution is to collect a small fraction of local data samples. As shown in Fig. 5b with 10 devices, each device’s uploading only 2% of 500 local data samples yields the higher accuracy and faster training convergence of FLD compared to Vanilla FL and FD. More general methods of seed sample collection will be presented in Sec. 4.
4 Surrogate Data Exchange
This section focuses on two problems. One is to rectify nonIID datasets of devices (see Q6). NonIID data degrades the gains from model parameter and model output exchange methods, since their benefits are rooted in exploiting the diversity and correlation of their training data. The other issue is to complete training in a single communication round (Q7). Minimizing communication rounds is useful not only for reducing communication bandwidth and energy, but also for mitigating the risk of privacy leakages incurred during information exchanges. To address these problems, we introduce two surrogate methods that are equivalent to exchanging raw data samples in a communicationefficient and privacypreserving manner: by exchanging a summary of raw datasets (S6), or by collectively training a realistic synthetic sample generator (S7).
4.1 Federated Data Summarization
(S6) Data summarization extracts useful information from local datasets at each device. This can be utilized for rectifying nonIID training data. Furthermore, by aggregating the data summary at a single location, ML models can be trained without further exchanging raw data samples. Compared to raw datasets, their data summaries usually have much smaller sizes and less information about each sample, thereby saving communication bandwidth while preserving privacy. There are three broad categories of data summarization techniques [13].
1) Statistical Summary
: Basic forms of statistical summary are representative values, such as mean, median, sum, and variance. These values can be computed for every regular time interval, spatial region, or data class. More advanced forms of statistical summary include histograms and statistical models representing raw datasets. For histograms, as an example, bagofwords ML is based on the occurrence of different words, which is represented as a histogram. For statistical model approaches, the Gaussian mixture model (GMM) is often utilized in acoustic recognition for determining which class each acoustic signal segment belongs to.
2) Dimensionality Reduction
: As exemplified in Sec. 3.1, raw data dimension can be reduced by quantizing states in reinforcement learning or grouping data samples into labels for classification tasks. More generally, it is useful to project highdimensional raw data samples onto a lowdimensional space. In this direction, principle component analysis (PCA) is one popular way, representing each raw sample using a linear combination of orthogonal principle components. For nonlinear dimensionality reduction, feature extraction is another wellknown method. This step can be separately performed prior to ML, or integrated into a deep neural network whose low layers play the feature extracting role.
3) Data Subsampling: Selecting a small amount of samples from the dataset also plays a role of data summarization. A straightforward approach is random sampling, which however does not provide any guarantee on how well the sampled subset represents the original dataset. Alternatively, coreset [14] is a notable data subsampling technique, which provides theoretical error bounds when an ML model is trained using the coreset. For a stricter privacy requirement, there also exist methods of constructing coresets that preserve the differential privacy of data.
The above three categories are not necessarily independent of each other, but can be integrated into a single data summarization scheme. For instance, one can compute the coreset on features extracted from images, which is a combination of dimensionality reduction and data subsampling.
4.2 MultiHop Federated Data Augmentation
When devices need to exchange big data, it is more communication efficient to make them collectively train a data sample generator in a federated fashion. By sharing the sample generator, each device can locally generate synthetic data samples resembling the data samples stored in the other devices, without further communication.
(S7) Federated data augmentation (FAug) realizes this idea for an image classification task [9]
. Namely, in FAug (see Fig. 6a), devices collectively train a conditional generative adversarial network (GAN) stored at a server. The GAN consists of a discriminator neural network paired with another generator neural network to be downloaded by each device. These two neural networks are trained in a semisupervised learning way, in which the discriminator evaluates the authenticity of the synthetic samples produced by the generator. To train the GAN, every device should upload: (i) a vector specifying its
target labels of lacking data samples; and (ii) seed samples appended to each target label.These target label information and seed samples determine the communication payload sizes. Furthermore, they are often privacy sensitive, e.g., medical Xray images (samples) are private, and the medical checkup items (target labels) may easily reveal diagnosis results. Multihop federated data augmentation with sample compression (MultFAug) [15] preserves such target label and sample privacy while maximizing communication efficiency as detailed next.
1) MultiHop Communication: Target labels out of all labels are identified in a vector, referred to as sample distribution information (SDI). To preserve the target label privacy, the truthful SDI of each device, termed private SDI, should be hidden from others. To this end, each device inserts dummy label indicators into the private SDI, yielding its public SDI. For example, in Fig. 6a has the target label of lacking samples. Then, is its private SDI for labels . The public SDI is when a dummy label indicator is inserted into the label .
Multihop communication makes the public SDI of devices be more truthful. In Fig. 6a, consider the next hop of is given as having its target label . The private SDI of can identically become the public SDI, by hiding its privacy in the preceding public SDI of . Constructing truthful public SDI decreases dummy label indicators appended with redundant seed samples, thereby reducing the communication payload sizes. Furthermore, multihop communication can also provide path loss gains, particularly for faraway devices that cannot reach directly to the server.
2) Sample Compression: Every device under MultFAug compresses its seed samples by randomly removing a fraction of each sample. This sample compression reduces its privacy leakage. Indeed, more compressed samples become less similar to each other, when evaluating the similarity via the classical multidimensional scaling (MDS) algorithm. Moreover, the sample compression decreases communication payload sizes. This is viable by converting the compressed samples into compressed sparse row (CSR) formats. Each CSR representation only describes its original sample’s nonzero elements with their row and column indices, thereby reducing the communication overhead.
The effectiveness of MultFAug on Vanilla FL under nonIID data is shown in Fig. 6b. For data privacy, the more hops, the better, allowing devices to hide their privacy more using less dummy labels, i.e., perhop payload size reduction. For test accuracy, there exists an optimal number of hops. Indeed, more hops entail more transmission attempts, balanced by the perhop payload reductions and path loss gains. This highlights the importance of network architecture and communication protocol designs, under the tradeoff between privacy and communicationefficiency.
5 Discussion and Conclusions
We conclude this article with a brief discussion about extending our proposed FML designs.
1) Realistic RAN Characteristics: Wireless interference and channel fluctuations may perturb information exchanges, inserting undesired noise into ML operations. To address this, realistic channel dynamics with antenna patterns and their impact on interference should be taken into account. Furthermore, different types of exchanges may have distinct latency and reliability requirements. Selecting their suitable RAN service links and optimally multiplexing them are important challenges. Lastly, FML enables ML aided fog RAN, from which FML can also benefit. Therefore, it is necessary to jointly optimize FML and fog RAN operations.
2) Heterogeneous Hardware Specifications: The available energy, computing power, and memory size of every device are different from each other, and an order of magnitude smaller than the edge server’s. Adaptive FL needs to account for these heterogeneous hardware specifications for optimizing local computing iterations and global communication rounds. Particularly under heterogeneous memory capacities, FD becomes more useful, which allows devices having different model sizes to collaborate with each other, as opposed to model parameter exchanging methods. Lastly, FLD and FAug transform the server’s high computing power into the solutions to devices’ limited transmission power and nonIID datasets, respectively. Depending on the computing power difference between the edge server and devices, optimizing the server’s role is an interesting problem.
3) User Privacy and Fairness: Usergenerated data can preserve its privacy by distorting raw samples or adding dummy information in FML. Therefore, preserving more privacy without losing accuracy requires more federating devices with more communication rounds in a less energyefficient way. In this respect, balancing privacy against communication and energy costs, the aforementioned 1) and 2) should be codesigned. Finally, preceding FML frameworks serve devices in a besteffort way, and equally aggregate local updates. The resulting global information can be biased, since it more frequently reflects the local datasets of nonstraggling devices having good channel conditions and high computing power. Possible solutions to this problem could fairly schedule the devices straggled in consecutive communication rounds, and could minimize a fairnessaware loss function, rather than the current empirical risk minimization.
References
 [1] J. Park, S. Samarakoon, M. Bennis, and M. Debbah, “Wireless network intelligence at the edge,” submitted to Proceedings of the IEEE [Online]. Arxiv preprint, vol. abs/1812.02858, December 2018.
 [2] H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communicationefficient learning of deep networks from decentralized data,” in Proc. of AISTATS, (Fort Lauderdale, FL, USA), Apr. 2017.
 [3] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated machine learning: Concept and applications,” ACM Trans. Intell. Syst. Technol., vol. 10, Feb. 2019.
 [4] P. Vepakomma, T. Swedish, R. Raskar, O. Gupta, and A. Dubey, “No peek: A survey of private distributed deep learning,” [Online]. ArXiv preprint: https://arxiv.org/abs/1812.03288.
 [5] S. Wang, T. Tuor, T. Salonidis, K. K. Leung, C. Makaya, T. He, and K. Chan, “Adaptive federated learning in resource constrained edge computing systems,” IEEE Journal on Selected Areas in Communications, vol. 37, pp. 1205–1221, June 2019.
 [6] A. Elgabli, J. Park, A. S. Bedi, and V. Aggarwal, “GADMM: Fast and communication efficient framework for distributed machine learning,” submitted to 2019 Neural Information Processing Systems (NeurIPS).
 [7] H. Kim, J. Park, M. Bennis, and S.L. Kim, “Blockchained ondevice federated learning,” to appear in IEEE Communications Letters [Online]. ArXiv preprint: abs/1808.03949.

[8]
P. Jiang and G. Agrawal, “A linear speedup analysis of distributed deep learning with sparse and quantized communication,” in
Advances in Neural Information Processing Systems 31 (S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. CesaBianchi, and R. Garnett, eds.), pp. 2525–2536, Curran Associates, Inc., 2018.  [9] E. Jeong, S. Oh, H. Kim, J. Park, M. Bennis, and S.L. Kim, “Communicationefficient ondevice machine learning: Federated distillation and augmentation under nonIID private data,” presented at Neural Information Processing Systems (NeurIPS) Wksp. Machine Learning on the Phone and other Consumer Devices (MLPCD), Dec. 2018. [online], ArXiv preprint, https://arxiv.org/abs/1811.11479.
 [10] R. Anil, G. Pereyra, A. Passos, R. Ormandi, G. E. Dahl, and G. E. Hinton, “Large scale distributed neural network training through online distillation,” ArXiv preprint, vol. abs//1804.03235, Apr. 2018.

[11]
H. Cha, J. Park, H. Kim, B. Bennis, and S.L. Kim, “Federated reinforcement
distillation with proxy experience memory,”
to be presented at 2019 International Joint Conference on Artificial Intelligence (IJCAI) Wksp. Federated Machine Learning for User Privacy and Data Confidentiality (FML)
.  [12] J. Park, S.L. Kim, and J. Zander, “Tractable resource management with uplink decoupled millimeterwave overlay in ultradense cellular networks,” IEEE Transactions on Wireless Communications, vol. 15, pp. 4362–4379, June 2016.
 [13] B. Ko, S. Wang, T. He, and D. ConwayJones, “On data summarization for machine learning in multiorganization federations,” in in Proc. of Workshop on Distributed Analytics InfraStructure and Algorithms for MultiOrganization Federations (DAIS), 2019.
 [14] H. Lu, M.J. Li, T. He, S. Wang, V. Narayanan, and K. S. Chan, “Robust coreset construction for distributed machine learning,” [Online]. ArXiv preprint: https://arxiv.org/abs/1904.05961.
 [15] E. Jeong, S. Oh, J. Park, H. Kim, B. Bennis, and S.L. Kim, “Multihop federated private data augmentation with sample compression,” to be presented at 2019 International Joint Conference on Artificial Intelligence (IJCAI) Wksp. Federated Machine Learning for User Privacy and Data Confidentiality (FML).
Comments
There are no comments yet.